AI Hallucination

What is an AI hallucination?

In artificial intelligence, a hallucination is a response generated by an AI system that presents false or misleading information as fact. The term is a loose analogy to human perception, even though language models have neither senses nor consciousness. What makes hallucinations tricky is that the invented statement is usually fluent, plausible and confidently phrased — an invented quotation, a non-existent citation or a wrong date. Well-known cases include chatbots inventing academic references or court rulings; the support chatbot of the airline Air Canada once cited a refund policy that did not exist.

Why do language models hallucinate?

Large language models (LLMs) are trained to predict the most probable next word. They hold no internal model of truth; they continue text in a way that is statistically consistent with their training data. When specific knowledge is missing, the model still "guesses" a plausible continuation instead of admitting uncertainty. OpenAI researchers have argued that hallucinations persist partly because the training and evaluation of LLMs reward guessing over acknowledging uncertainty. In practice, a model has no reliable signal for what it does not know, so it tends to produce an answer for almost any prompt.

Is "confabulation" the better term?

Some researchers consider "hallucination" misleading because it implies sensory perception. They prefer the psychiatric term confabulation — generating fabricated details that are not recognised as false. It better captures that these errors arise from gaps in information recall rather than from faulty perception.

How can the risk be reduced?

Hallucinations cannot be eliminated entirely, but they can be reduced. A common approach is Retrieval-Augmented Generation (RAG), where the model is given verified source texts and grounds its answer in them. Further measures include citing sources, writing precise prompts, requesting references, and independently verifying important facts. Breaking a task into smaller, checkable steps and cross-checking against a second source also lowers the error rate. For learning, the rule is simple: AI output is a starting point, not a trusted source — fact-checking remains a human task.